AI in customer support: what the 5,172-agent study measured

In short: most figures on AI and customer support productivity come from vendors. One of the few large independent field studies is Generative AI at Work by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published in the Quarterly Journal of Economics in 2025 after a 2023 NBER working paper. Following the staggered rollout of a generative AI conversational assistant among 5,172 customer support agents, it measures a 15% average increase in issues resolved per hour — with large differences between agents.

Update, September 23, 2026: the figures now come from the published version of the study (5,172 agents, +15%), not from the 2023 working paper (5,179 agents, +14%).

What the study measured

What it means for your support team

• →Measure by seniority, not only on average: in the study, most of the gain goes to newer or less skilled agents.
• →Treat the assistant as a way to spread the answers of your best agents, and keep an eye on quality among experienced ones, where the study found a slight decline.
• →Look at rare cases: the gains are largest on problems agents meet too seldom to master, provided the system has enough examples of them.
• →Track escalations and the tone of conversations alongside speed: the study saw fewer requests for a manager.

A bot that answers customers directly is a different case

The study measured an assistant helping human agents, not a bot replacing them in front of the customer. For a bot that talks to customers, two points change. First, since August 2, 2026, article 50 of the AI Act requires that people interacting with an AI system be told so, unless it is obvious from the context (see our article on the AI Act in France). Second, disclosure has a cost: in Luo, Tong, Fang and Qu (Marketing Science, 2019), revealing a chatbot’s identity reduced customer purchases, because customers perceived the disclosed bot as less knowledgeable and less empathetic. The answer is not to hide the bot — the obligation is clear — but to limit it to what it handles well and to hand over to a person quickly.

On WhatsApp, every reply now has a price: see what becomes paid on the WhatsApp Business Platform. A conversation resolved in two messages costs less than one that loops.